TechWithAdit (TWA)

Three model launches are converging today. SpaceXAI and Cursor are dropping their first jointly built AI model right now — trained on Colossus, shipped inside the world's most popular coding tool. OpenAI just got US government clearance to do a broad rollout of GPT-5.6 this week. And in the background, DeepSeek has quietly started building its own AI chip — sending Nvidia down 1.6% in premarket.

On the legal front, British Columbia is suing OpenAI over a school shooting it says the company could have prevented — and Google Cloud is making a sovereignty push to host its latest AI models directly inside India.

According to an internal staff memo seen by The Information, SpaceXAI and Cursor plan to drop their first jointly developed AI model as early as today. The launch was pushed back earlier this week to improve efficiency. The model was trained on Colossus — SpaceX's supercomputer with roughly one million H100-equivalent GPUs — and ships inside both the Cursor editor and Grok Build. It is the first product to emerge from SpaceX's $60 billion acquisition of Anysphere (Cursor's maker), and the first SpaceXAI model built on Cursor's real production code from millions of professional developers.

This matters because Cursor was the last major model-neutral coding tool at scale — developers could pick Claude, GPT, or any frontier model. With SpaceXAI taking over, that neutrality ends. The data flywheel is now live: every line of code written in Cursor trains the SpaceXAI model, which improves Cursor, which attracts more engineers. xAI, which became SpaceXAI on Monday, is now building a closed vertical stack across compute, model, data, and distribution — in one product.

The insights: For developers, the key question is whether Cursor's model-neutral era is truly over. For enterprises, it's a code custody question — your engineers' production code is now feeding a competitor's model. The race to own the developer layer of AI just got its most consequential move yet.

Reuters reported Tuesday that Chinese AI startup DeepSeek has begun developing its own inference chip, according to three people familiar with the matter. The project has been underway for about a year and is in early stages — DeepSeek is seeking external partners in chip design, foundry, and memory. The chip targets inference workloads (running models, not training them), where demand is growing fastest as AI apps scale. Nvidia shares slipped 1.6% in premarket on the news.

The strategic logic is clear. US export controls have barred Chinese companies from buying Nvidia's most advanced chips. Huawei has filled part of the gap, but dependency on any external supplier carries risk. DeepSeek joins OpenAI (Jalapeño), Anthropic (Samsung talks), and Chinese rivals Alibaba and Baidu in pursuing custom silicon. An analyst at Radio Free Mobile noted: "Nvidia is at zero in China and staying there." But the pressure to build domestic alternatives is now systemic, not just strategic.

The insights: Every major AI company is now trying to own its chip stack. DeepSeek entering the race confirms the pattern: if you're a frontier AI lab with $50B+ valuation and a geopolitical reason to reduce dependency, you build your own silicon. The inference chip market — not training — is where the next hardware war plays out.

Google Cloud CEO Thomas Kurian has announced that Google will now host its latest AI models directly in India, a significant step beyond having data centre infrastructure in the country. The move means Indian enterprises and government bodies can run Google's most advanced Gemini models on Indian soil, with data residency guarantees — addressing the sovereignty concerns that have held back large regulated sectors from fully committing to cloud AI. This builds on Google's $15 billion Vizag AI Hub commitment and positions India-based cloud as a first-class deployment option for frontier AI, not just a secondary region.

The insights: Data sovereignty is no longer a compliance checkbox for Indian enterprises — it's a procurement requirement. Google hosting its latest models in-country changes the conversation with banking, healthcare, and government clients who couldn't move to cloud AI before. For Indian operators, this is the clearest signal yet that frontier AI is now available without the sovereignty trade-off.

British Columbia's provincial government announced it is preparing a lawsuit against OpenAI for failing to alert law enforcement to violent ChatGPT activity before the Tumbler Ridge school shooting in February 2026, in which 18-year-old Jesse Van Rootselaar killed eight people including five children and a teacher. OpenAI had banned Van Rootselaar's account in June 2025 — months before the attack — but did not report the violent prompts to police. The province's Attorney General Niki Sharma said BC wants to "hold OpenAI and its decision-makers accountable." OpenAI CEO Sam Altman apologized in April but said the account was suspended without evidence of an imminent threat. Lawyers for the families allege OpenAI stayed silent because "reporting one case would mean reporting thousands."

The insights: This case will define whether AI companies have a duty to report dangerous user behaviour to law enforcement — even when no imminent threat is explicitly stated. The answer from BC's court system will set a precedent that every AI platform operating in the English-speaking world is watching closely.

Operator's Spotlight Read

The US Department of Commerce has approved a broad launch of GPT-5.6, Axios reported Tuesday, citing a source familiar with the matter. OpenAI expects to roll out GPT-5.6 widely this week. Testing was conducted by the Commerce Department's Center for AI Standards and Innovation, with OpenAI sending technical experts to Washington to address any questions. Since June 26, GPT-5.6 had been restricted to roughly 20 government-approved partner organisations — a staggered release OpenAI explicitly said was "not our preferred way to release new models" but agreed to at the government's request.

The GPT-5.6 family consists of three tiers — Sol (flagship, new state-of-the-art on Terminal-Bench 2.1), Terra (GPT-5.5 performance at half the cost), and Luna (fastest and cheapest). Sol is specifically noted for advances in cyber vulnerability research and exploitation — which is exactly why it triggered the government review process in the first place. The clearance follows the same framework used to restore Anthropic's Fable 5 and Mythos on June 30. A repeatable process is now forming: lab previews the model with the government, testing happens, clearance is granted, broad rollout follows.

The insights: The US government has established a de facto pre-clearance regime for frontier AI models. This is no longer a one-off — it's the new normal. Every major model release from here will go through this process before reaching the public. For operators building on GPT-5.6 or any future frontier model, the variable is no longer just capability. It's how long the government review takes, and whether your use case is one the reviewers are comfortable with.

Was this email forwarded to you? Don't miss any updates — Subscribe to TechWithAdit for sharp, no-noise tech intelligence. Stay sharp. — Adit